0.1.47

Init-only runs that exit with a non-zero code now surface a more actionable
message naming the exit code and the likely cause (unsupported model or
rejected session) instead of the generic "did not produce a result" text.
Helps operators diagnose model-id / billing-tier failures (e.g. opus 4.6).

Co-Authored-By: Paperclip <noreply@paperclip.ing>

package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "paperclip-adapter-claude-k8s",
-  "version": "0.1.45",
+  "version": "0.1.47",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "paperclip-adapter-claude-k8s",
-      "version": "0.1.45",
+      "version": "0.1.47",
       "license": "MIT",
       "dependencies": {
         "@kubernetes/client-node": "^1.0.0",

package.json

@@ -1,6 +1,6 @@
 {
   "name": "paperclip-adapter-claude-k8s",
-  "version": "0.1.45",
+  "version": "0.1.47",
   "description": "Paperclip adapter plugin that runs Claude Code agents as Kubernetes Jobs",
   "license": "MIT",
   "repository": {

src/server/execute.test.ts

@@ -60,7 +60,7 @@ vi.mock("@paperclipai/adapter-utils/server-utils", async (importOriginal) => {
   });
 });
 
-const { isK8s404, buildPartialRunError, classifyOrphan, describePodTerminatedError, streamPodLogsOnce, shouldAbortForCancellation, execute } = await import("./execute.js");
+const { isK8s404, buildPartialRunError, classifyOrphan, describePodTerminatedError, describeTruncationCause, streamPodLogsOnce, shouldAbortForCancellation, execute } = await import("./execute.js");
 
 function makeJob(opts: {
   runId?: string;
@@ -150,10 +150,10 @@ describe("buildPartialRunError", () => {
     expect(buildPartialRunError(null, "", "")).toBe("Claude exited with code -1");
   });
 
-  it("skips system/init events and returns generic message when only init captured", () => {
+  it("returns init-only message when stdout is init-only with non-zero exit code (FAR-101)", () => {
     const msg = buildPartialRunError(1, "claude-sonnet-4-6", initLine);
     expect(msg).toBe(
-      "Claude started but did not produce a result (model: claude-sonnet-4-6) — check API credentials, model support, and adapter config",
+      "Claude exited immediately after init (model: claude-sonnet-4-6) (exit code 1) — the model may be unsupported or the session may have been rejected before producing output",
     );
   });
 
@@ -170,15 +170,15 @@ describe("buildPartialRunError", () => {
     expect(msg).toBe("Claude exited with code 1: Error: no API key configured");
   });
 
-  it("skips result events (structured protocol artefact — not surfaced verbatim)", () => {
+  it("returns init-only message when stdout has init + result event but no plain content (structured artefact, not surfaced verbatim)", () => {
     // In production, buildPartialRunError is only called when parseClaudeStreamJson
     // returns null (no result event).  If somehow a result event appears here, the
-    // raw JSON blob must not be shown — the "did not produce a result" message is
+    // raw JSON blob must not be shown — the init-only message is cleaner and avoids
-    // cleaner and avoids leaking protocol internals to the UI.
+    // leaking protocol internals to the UI.
     const resultLike = JSON.stringify({ type: "result", subtype: "error", result: "rate limit" });
     const stdout = [initLine, resultLike].join("\n");
     const msg = buildPartialRunError(2, "claude-sonnet-4-6", stdout);
-    expect(msg).toContain("did not produce a result");
+    expect(msg).toContain("Claude exited immediately after init");
     expect(msg).toContain("claude-sonnet-4-6");
     expect(msg).not.toMatch(/\{.*type.*result/);
   });
@@ -362,6 +362,33 @@ describe("describePodTerminatedError", () => {
   });
 });
 
+describe("describeTruncationCause", () => {
+  it("annotates exit code 137 as SIGKILL/OOM", () => {
+    const msg = describeTruncationCause({ exitCode: 137, reason: "OOMKilled", message: "Memory cgroup out of memory", signal: null });
+    expect(msg).toContain("exit code 137");
+    expect(msg).toContain("SIGKILL");
+    expect(msg).toContain("OOMKilled");
+    expect(msg).toContain("Memory cgroup out of memory");
+  });
+
+  it("annotates exit code 143 as SIGTERM", () => {
+    const msg = describeTruncationCause({ exitCode: 143, reason: null, message: null, signal: null });
+    expect(msg).toContain("exit code 143");
+    expect(msg).toContain("SIGTERM");
+  });
+
+  it("falls back to 'pod state unavailable' when state is null", () => {
+    const msg = describeTruncationCause(null);
+    expect(msg).toContain("pod state unavailable");
+  });
+
+  it("emits 'no exit code' when exitCode is null but state exists", () => {
+    const msg = describeTruncationCause({ exitCode: null, reason: "Error", message: null, signal: null });
+    expect(msg).toContain("no exit code");
+    expect(msg).toContain("reason=Error");
+  });
+});
+
 describe("execute: all-invalid agent.id (N4)", () => {
   it("returns hard error without creating a Job when agent.id sanitizes to null", async () => {
     const logs: string[] = [];
@@ -1019,7 +1046,7 @@ describe("execute: happy path", () => {
       },
     );
     mockCoreListPods.mockResolvedValue({
-      items: [{ metadata: { name: "pod-abc" }, status: { containerStatuses: [{ name: "claude", state: { terminated: { exitCode: 137 } } }] } }],
+      items: [{ metadata: { name: "pod-abc" }, status: { containerStatuses: [{ name: "claude", state: { terminated: { exitCode: 137, reason: "OOMKilled", message: "Memory cgroup out of memory" } } }] } }],
     });
 
     const executePromise = execute(makeCtx());
@@ -1030,6 +1057,9 @@ describe("execute: happy path", () => {
     expect(result.errorMessage).toContain("truncated mid-stream");
     expect(result.errorMessage).toContain("claude-opus-4-7");
     expect(result.errorMessage).toContain("exit code 137");
+    expect(result.errorMessage).toContain("SIGKILL");
+    expect(result.errorMessage).toContain("OOMKilled");
+    expect(result.errorMessage).toContain("Memory cgroup out of memory");
   });
 
   it("reconnects log stream and logs status when job completion takes > 3s", async () => {

src/server/execute.ts

@@ -110,24 +110,12 @@ export function shouldAbortForCancellation(runStatus: string | undefined): boole
 }
 
 /**
- * Build the error message when Claude's stdout contains no result event.
+ * Returns the first non-JSON/plain-text line in stdout, treating JSON objects
- * Skips system/init event lines so the UI doesn't display the raw init JSON.
+ * with a "type" field as protocol artefacts and skipping them.
- * Exported for unit tests.
+ * Used by buildPartialRunError to detect init-only runs.
  */
-export function buildPartialRunError(
+function firstContentLine(stdout: string): string {
-  exitCode: number | null,
+  return stdout.split(/\r?\n/)
-  model: string,
-  stdout: string,
-): string {
-  if (exitCode === 0) return "Failed to parse Claude JSON output";
-
-  // Walk stdout lines and skip every structured streaming event (any JSON
-  // object that carries a non-empty "type" field: system, assistant, user,
-  // rate_limit_event, result, …).  All of these are protocol artefacts and
-  // produce confusing raw-JSON blobs when surfaced verbatim as an error
-  // message.  Only plain-text lines (non-JSON, or JSON without a type field)
-  // are treated as human-readable content worth including in the error.
-  const firstContentLine = stdout.split(/\r?\n/)
     .map((l) => l.trim())
     .find((l) => {
       if (!l) return false;
@@ -142,19 +130,55 @@ export function buildPartialRunError(
       }
       return true;
     }) ?? "";
+}
+
+/**
+ * Returns true when stdout contains only init/system/assistant events from the
+ * given model with no human-readable content lines.  Used to detect init-only
+ * non-zero-exit runs that should be classified as claude_init_failed rather than
+ * the generic "Claude exited with code N" message.
+ */
+function isInitOnlyRun(model: string, stdout: string): boolean {
+  if (!stdout.trim() || !model) return false;
+  const content = firstContentLine(stdout);
+  if (content) return false;
+  // Check that at least the init event for this model was seen
+  const hasModelInit = stdout.includes(`"model":"${model}"`) || stdout.includes(`"model":"${model.replace(/-/g, "_")}"`);
+  return hasModelInit;
+}
+
+/**
+ * Build the error message when Claude's stdout contains no result event.
+ * Skips system/init event lines so the UI doesn't display the raw init JSON.
+ * Exported for unit tests.
+ */
+export function buildPartialRunError(
+  exitCode: number | null,
+  model: string,
+  stdout: string,
+): string {
+  if (exitCode === 0) return "Failed to parse Claude JSON output";
 
   // If the stream contained only structured events with no plain-text output,
   // surface the model name so the operator can diagnose missing credentials
   // or unsupported/misconfigured model.
-  const initOnlyOutput = stdout.trim() !== "" && model !== "" && !firstContentLine;
+  const contentLine = firstContentLine(stdout);
+  if (contentLine) {
+    return `Claude exited with code ${exitCode ?? -1}: ${contentLine}`;
+  }
+
+  if (isInitOnlyRun(model, stdout) && (exitCode ?? 0) !== 0) {
+    const modelHint = model ? ` (model: ${model})` : "";
+    return `Claude exited immediately after init${modelHint} (exit code ${exitCode ?? -1}) — the model may be unsupported or the session may have been rejected before producing output`;
+  }
+
+  const initOnlyOutput = stdout.trim() !== "" && model !== "";
   if (initOnlyOutput) {
     const modelHint = model ? ` (model: ${model})` : "";
     return `Claude started but did not produce a result${modelHint} — check API credentials, model support, and adapter config`;
   }
 
-  return firstContentLine
+  return `Claude exited with code ${exitCode ?? -1}`;
-    ? `Claude exited with code ${exitCode ?? -1}: ${firstContentLine}`
-    : `Claude exited with code ${exitCode ?? -1}`;
 }
 
 export type OrphanClassification =
@@ -574,6 +598,27 @@ async function waitForJobCompletion(
  * Get the exit code from the Job's pod.
  */
 async function getPodExitCode(namespace: string, jobName: string, kubeconfigPath?: string): Promise<number | null> {
+  const state = await getPodTerminatedState(namespace, jobName, kubeconfigPath);
+  return state?.exitCode ?? null;
+}
+
+/**
+ * Get the claude container's terminated state (exit code, reason, message,
+ * signal) from the Job's pod. Returns null if the pod or container is gone.
+ * Used by the no-result error path to explain *why* a run was truncated.
+ */
+export interface PodTerminatedState {
+  exitCode: number | null;
+  reason: string | null;
+  message: string | null;
+  signal: number | null;
+}
+
+async function getPodTerminatedState(
+  namespace: string,
+  jobName: string,
+  kubeconfigPath?: string,
+): Promise<PodTerminatedState | null> {
   const coreApi = getCoreApi(kubeconfigPath);
   const podList = await coreApi.listNamespacedPod({
     namespace,
@@ -583,7 +628,40 @@ async function getPodExitCode(namespace: string, jobName: string, kubeconfigPath
   if (!pod) return null;
 
   const containerStatus = pod.status?.containerStatuses?.find((s) => s.name === "claude");
-  return containerStatus?.state?.terminated?.exitCode ?? null;
+  const terminated = containerStatus?.state?.terminated;
+  if (!terminated) return null;
+  return {
+    exitCode: terminated.exitCode ?? null,
+    reason: terminated.reason ?? null,
+    message: (terminated.message ?? "").trim() || null,
+    signal: terminated.signal ?? null,
+  };
+}
+
+/**
+ * Format a human-readable explanation for a truncated run, including the
+ * pod's claude-container terminated state when available. Exit code 137
+ * is annotated as SIGKILL/OOM since that is the most common cause.
+ * Exported for unit tests.
+ */
+export function describeTruncationCause(
+  state: PodTerminatedState | null,
+): string {
+  if (!state) {
+    return "pod state unavailable — likely deleted before exit could be read";
+  }
+  const parts: string[] = [];
+  if (state.exitCode !== null) {
+    parts.push(`exit code ${state.exitCode}`);
+    if (state.exitCode === 137) parts.push("SIGKILL (commonly OOMKilled)");
+    else if (state.exitCode === 143) parts.push("SIGTERM");
+  } else {
+    parts.push("no exit code");
+  }
+  if (state.signal !== null) parts.push(`signal ${state.signal}`);
+  if (state.reason) parts.push(`reason=${state.reason}`);
+  if (state.message) parts.push(`message=${state.message}`);
+  return parts.join(", ");
 }
 
 /**
@@ -998,6 +1076,7 @@ export async function execute(ctx: AdapterExecutionContext): Promise<AdapterExec
 
   let stdout = "";
   let exitCode: number | null = null;
+  let podTerminatedState: PodTerminatedState | null = null;
   let jobTimedOut = false;
   let keepaliveTimer: ReturnType<typeof setInterval> | null = null;
   // Set when we return a mismatch error so the finally block knows not to
@@ -1297,7 +1376,8 @@ export async function execute(ctx: AdapterExecutionContext): Promise<AdapterExec
       }
     }
 
-    exitCode = await getPodExitCode(namespace, jobName, kubeconfigPath);
+    podTerminatedState = await getPodTerminatedState(namespace, jobName, kubeconfigPath);
+    exitCode = podTerminatedState?.exitCode ?? null;
   } finally {
     if (keepaliveTimer) clearInterval(keepaliveTimer);
     activeJobs.delete(activeJobRef);
@@ -1368,13 +1448,13 @@ export async function execute(ctx: AdapterExecutionContext): Promise<AdapterExec
       };
     }
     if (parsedStream.truncatedMidStream) {
-      const exitHint = exitCode === null ? "no exit code" : `exit code ${exitCode}`;
+      const cause = describeTruncationCause(podTerminatedState);
       const modelHint = parsedStream.model ? ` (model: ${parsedStream.model})` : "";
       return {
         exitCode,
         signal: null,
         timedOut: false,
-        errorMessage: `Claude run was truncated mid-stream${modelHint} — assistant produced content but no result event arrived (${exitHint}); pod may have been terminated, OOMKilled, or the CLI crashed`,
+        errorMessage: `Claude run was truncated mid-stream${modelHint} — assistant produced content but no result event arrived; ${cause}`,
         errorCode: "claude_truncated",
         resultJson: { stdout },
       };